Purdue team among three finalists for prestigious supercomputing award

A team led by Purdue electrical and computer engineering professors has been named one of the three finalists for the Association of Computing Machinery’s Gordon Bell Prize for their work in image reconstruction. The winner of the prize will be announced at the upcoming SC17 supercomputing conference, which will be held in Denver from Nov. 12-17.

Purdue's new Brown community cluster should make the TOP500 list of the world's most powerful supercomputers at SC17. Meanwhile, ITaP staff will be keeping busy running the conference's network, presenting in a variety of workshops about high-performance computing and more.

When you go to the doctor for a computed tomography (CT) scan, the machine doesn’t take a simple picture, the way an X-ray machine does. Rather, it generates many different pieces of data that must be reconstructed, or assembled like a puzzle, into the final image your doctor interprets. Purdue professors Charles Bouman and Samuel Midkiff work on developing algorithms to make those final images more accurate, or equally as accurate with less data, that is, less exposure to radiation for the patient.

In particular, they focus on a category of image reconstruction algorithms known as Model-Based Iterative Reconstruction (MBIR) algorithms, which produce more accurate images than their faster counterparts, but have in the past generally been considered too slow for use in many applications. Their selection as Gordon Bell Prize finalists is based on a new type of MBIR algorithm they’ve developed that is much faster than its predecessors, especially when run on a supercomputer with thousands of processor cores.

The new algorithm developed by Bouman and Midkiff along with former graduate students Xiao Wang, Amit Sabne and Putt Sakdhnagool, and Sherman Kisner of High Performance Imaging LLC, runs about 18 times faster than previous state-of-the-art algorithms on a single core. But because it’s specifically designed to take advantage of a supercomputer’s parallel processing ability, when the comparison takes place on a supercomputer, their algorithm beats existing algorithms by a much larger margin – running approximately 1,600 times faster than the current state-of-the-art.

A comparison showing the improvement in reconstruction with the Purdue team's algorithm (R) versus an existing algorithm (L) on an X-ray diffraction image. A comparison showing the improvement in reconstruction with the Purdue team's algorithm (R) versus an existing algorithm (L) on an X-ray diffraction image.

“Generally, you’re happy if you get a speed-up of two or three times, and for a lot of things, you’re happy if you get a speed-up of 10 percent or more,” says Midkiff, a professor of electrical and computer engineering. “A speed-up of over 1,000 is very dramatic.”

“It was sort of widely believed this problem couldn’t be mapped to really big computers,” says Bouman, the Showalter Professor of Electrical and Computer Engineering and Biomedical Engineering. “By taking the same problem and flipping it upside down and reorganizing the data and the operations, we were able to show that we can scale it up to computers with thousands of cores.”

As Gordon Bell Prize finalists, the Purdue team will deliver a presentation about their work at SC17 at 11:30 a.m. MST on Thursday, Nov. 16. They will also be speaking in the Purdue booth, #1571, at 10 a.m. MST on Wednesday, Nov. 15.

The award is named after and funded by support from high-performance computing pioneer and Microsoft researcher emeritus Gordon Bell, who will also speak at SC17. It is awarded annually for outstanding achievement in high-performance computing, with particular emphasis on applying high-performance computing to applications in science, engineering and large-scale data analytics.

The development of Bouman and Midkiff’s algorithm was largely performed on ITaP’s Rice community cluster supercomputer. The team has also developed an even faster version optimized for a graphical processing unit (GPU) and their next steps include expanding that algorithm to run on a multi-core GPU, something they’ll do with the help of the Halstead-GPU cluster.

Although medical CT scans are probably the most familiar kind of tomography, the same technology has broad applications to many different disciplines of science and engineering as well as daily life. Everything from particle accelerators to the scanners that review the contents of travelers’ checked bags at the airport use tomography and can benefit from faster and more efficient reconstruction algorithms such as the one developed by the Purdue team.

“For a lot of applications, this method will be much more practical or much cheaper for people to use,” says Midkiff.

Bouman and Midkiff have co-founded a start-up, High Performance Imaging, through the Purdue Foundry, and are working to commercialize their technology. They expect to have customers purchasing this new technology within the next year.

Writer: Adrienne Miller, science and technology writer, Information Technology at Purdue (ITaP), 765-496-8204, mill2027@purdue.edu.

Last updated: November 10, 2017

Purdue University, 610 Purdue Mall, West Lafayette, IN 47907, (765) 494-4600

© 2015 Purdue University | An equal access/equal opportunity university | Copyright Complaints | Maintained by ITaP

Trouble with this page? Disability-related accessibility issue? Please contact ITaP at itap@purdue.edu.